One-Class Hybrid Heterogeneous Network for Detecting HEVC Double Compression With the Same Coding Parameters

High-efficiency video coding (HEVC) is a recent yet increasing widely used video coding standard, and double compression detection is usually an essential step to verify the integrity of HEVC-encoded videos. However, it is challenging due to fewer traces left by HEVC double compression with the same parameters. Moreover, existing full-supervised learning works for HEVC double compression detection are inefficient because they depend on large amounts of labeled pristine and forged videos, which are difficult to be collected. To address these issues, a one-class classification (OCC)-based hybrid heterogeneous network is proposed, which only needs the pristine videos. We first develop a modulation layer with both motion alignment and high-frequency preservation operations, which serves as an effective metric to evaluate the differences between those videos compressed once and twice. Then, a heterogeneous network with a shallow convolutional neural network (CNN) and a six-node graph neural network (GNN) is proposed. Specifically, the shallow CNN, which pays more attention to medium or fast-motion regions, learns from subtle fluctuations of pixel values caused by double compression, whereas GNN, which focuses on static and slow-motion regions, is developed to represent the local-global relationship of video patches and the distribution of zero-value pixels in the high-frequency components of motion-aligned residuals. Due to the motion-aware mechanism, the proposed approach only learns features from single compressed videos. Extensive experimental results show that the proposed approach outperforms the state-of-the-art full-supervised learning works and other more complex OCC works.